A motor drive must regulate torque and speed and respond to changing loads. On the same device, its manufacturer would like to recognise unusual vibration or an electrical current pattern that changes over time. These requirements increasingly coexist: control the machine and understand how it is operating, while keeping firmware timing predictable.
With the new STM32H7P family, announced on 4 November 2026, STMicroelectronics combines a processor with two accelerators for signals and artificial intelligence. For industrial electronics developers, the question is how to use them while keeping control timing predictable: which tasks to assign to each block, what data to collect and what timing to check on a prototype.
STM32H7P: what changes in the architecture
The ST announcement introduces a family targeting motor control, power conversion and machine condition monitoring. The STM32H7P4/P5 page summarises its main blocks. Frequencies, capacities and performance figures in the table are manufacturer claims.
Main blocks and their role in the design
| Resource | Role | What the project must verify |
|---|---|---|
| Cortex-M7 up to 520 MHz | Executes firmware | Deadlines, priorities and remaining CPU load. |
| HSP | Accelerates mathematics and signal processing | Available functions and transfer overhead. |
| Neural-ART NPU up to 85 GOPS | Accelerates AI inference | Model compatibility, memory and latency. |
| Flash up to 2 MB | Stores code and non-volatile data | Space for firmware, model and updates. |
HSP means Hardware Signal Processor; NPU means Neural Processing Unit. The first resource concerns signal processing, the second neural network calculations. 85 GOPS equals 0.085 TOPS, or billions of operations per second. That figure does not tell us how many milliseconds a particular product’s model will take: supported operations, numerical formats, memory and work remaining on the CPU all matter.
Motor control: timing comes first
In a motor drive, firmware acquires measurements, updates the controller and commands the power stage. One example is FOC, Field Oriented Control, included in ST’s motor control ecosystem. Current measurements and rotor position information allow quantities associated with flux and torque to be regulated.
The decisive constraint is the time available between a measurement and a command update. Take a 20 kHz loop as an example: its period would be 50 microseconds. That is the time available to finish the work scheduled for that cycle. Vibration analysis that considers tens or hundreds of milliseconds of signal follows a very different rhythm.
Inference should therefore not automatically enter the path that determines the motor command. An initial design can keep diagnostics as an observer: it receives valid data, produces an indicator and informs supervision. Any intervention on the machine needs defined logic, consistent with protection functions and safety requirements.
HSP: preparing the signal before looking for anomalies
A sensor produces a sequence of numbers; a model needs data that preserve useful information. DSP, Digital Signal Processing, covers the work between the two: filtering, offset removal, normalisation, frequency analysis and feature extraction.
For a motor, possible observations include vibration RMS amplitude, energy within selected bands or changes in spectral components. The FFT, Fast Fourier Transform, is an algorithm for computing the discrete Fourier transform and examining frequency content. It does not explain a peak’s cause on its own: that content must be related to speed, load and mechanical characteristics.
The HSP’s potential benefit is moving suitable operations from software execution to a dedicated resource. The benefit needs measurement including data preparation, setup, transfers and result retrieval. For a small operation, management overhead may be as significant as computation. The choice of functions to accelerate depends on the libraries and tool versions actually supported on the device.
NPU: what the model can tell us
Inference applies an already trained model to new data. A classifier might distinguish known conditions; an anomaly detector might indicate departure from learned behaviour. Here, Edge AI means performing that computation on the device, near the source of the data.
The NPU can accelerate compatible parts of the network. The complete path still includes acquisition, preparation, any CPU operations and interpretation of the output. A very fast model can be of little use if it receives incomplete windows or data unlike those used during development.
A change in vibration deserves attention, but it does not explain the cause on its own. It could come from a new load, structural resonance or a remounted sensor. Identifying a fault, and especially estimating how long remains before it occurs, requires specific data and validation. The indicator therefore needs to retain the context in which it was produced.
One possible division of tasks: the CPU handles control, the HSP processes signals and the NPU runs the model. The figure summarises the role of the three blocks.
An example: vibration in a motor drive
Imagine adding a vibration sensor to a motor drive. Firmware would need to collect samples, retain their sequence and operating conditions, prepare model input and make the result available to supervision. A few calculations can help show what this would involve before building a prototype.
Suppose acquisition runs at 6,400 samples/s per axis and analysis uses 1,024-sample windows. Each window lasts 1,024 / 6,400 = 0.16 s, or 160 ms. A decision based on a complete window already depends on this observation interval, before computation is added.
For an FFT of that length, spectrum points are spaced by 6,400 / 1,024 = 6.25 Hz. This is frequency bin spacing, rather than a guarantee of separating any two phenomena 6.25 Hz apart: windowing, noise and signal stability affect interpretation. A longer window changes the tradeoff between frequency observation, delay and memory.
To update the indicator more frequently, windows could advance by 256 samples: a new input every 40 ms, with 75% overlap and 25 evaluations per second. Each window still covers 160 ms. A higher update rate does not automatically shorten the analysed signal interval, and it increases the workload that must be processed.
The result should carry at least the window timestamp, data quality and operating state. A measurement during a speed ramp can mean something different from the same measurement at steady speed. Sensor type, useful bandwidth and mounting also need to match the phenomenon: collecting more samples cannot recover information the sensor did not acquire properly.
The example’s data path: collect the signal, split it into windows, extract features and pass them to the model. Any resulting alert would then need to be assessed alongside the machine’s operating state.
Memory and buffers: the model is only part of the budget
Continuing the example, three axes with samples stored at 16 bits need 3 × 1,024 × 2 = 6,144 bytes, or 6 KiB, for a raw window. Two buffers of that size need 12 KiB. The sample storage format would need to be chosen to suit the sensor and the planned processing.
Additional requirements include DSP buffers, model weights and activations, task stacks, communications and firmware diagnostics. The network file size is not the total memory needed during inference. Likewise, Flash does not replace the working memory required by the application.
Buffer management needs to define who may write and who may read. One possible arrangement alternates a buffer being filled and one being processed, but overlapping windows and a delayed consumer need a consistent policy. If analysis falls behind, the system must know whether to discard stale windows, reduce updates or report unavailable diagnostics. A growing queue produces increasingly old results.
Latency and jitter: what to measure
The diagnostic budget includes more than the neural network. An initial model is T_decision = T_observation + T_preparation + T_inference + T_alarm_rule. Waiting, transfers and result delivery must also be considered. If a rule requires several consecutive windows, the time to report an event changes again.
With a new input every 40 ms, the system must sustainably process that rate or apply an explicit load reduction strategy. For control, the primary concern is whether the extra activity causes a deadline violation. Latency is the delay along a path; jitter is its timing variation. Measuring only average inference time does not answer this question.
To check this, I would compare the same control with diagnostics off and on, adding communication traffic, interface requests and loads similar to those of the machine. I would measure the delay between acquisition and command, critical path duration and any missed deadlines. Pulses on a debug output or tracing tools would help observe these timings, while accounting for the disturbance introduced by measurement.
Appropriate priorities and accelerators help organise work, but shared resources, drivers and non-preemptible sections still need examination. Validation must also include timeouts, full buffers and late AI results. Acceptable behaviour should be decided before testing, rather than inferred after a favourable plot appears.
Two rhythms to accommodate: the periodic control cycle and the longer diagnostic operations. The prototype would need to show that diagnostics leave control enough time to finish its work.
Representative data, false alarms and unseen conditions
For a manufacturer, collecting the right data may be one of the hardest parts. Acquisitions should cover different machines, speeds and loads, including temperature changes, disturbances and mounting differences. Data from one machine during one shift can produce a model that recognises that session well and behaves differently in the field.
Training and validation should also be separated by machine or session when appropriate. Randomly splitting nearly identical windows, especially overlapping ones, risks placing very similar information in both sets and making evaluation too optimistic.
Machine users need to know how many false alarms to expect per shift, which anomalies might be missed and how long detection would take. A single accuracy percentage does a poor job of describing these questions. Thresholds and confirmation over multiple windows may reduce occasional alerts while adding delay or missing brief events.
Conditions outside validation coverage also need consideration: a disconnected sensor, saturated signal, new operating regime or mechanical modification. Firmware should distinguish unusable data from a valid diagnostic result. A missing result must not automatically become “everything is normal”.
Software tools, integration and updates
Before choosing tools, check their support for the exact microcontroller part number and the software versions the project will use. The official ST Edge AI tools overview presents tools for preparing and assessing models. The project would need to check supported operators and numerical formats, along with which parts of the network would run on the NPU or CPU.
Before connecting the model to control, compare identical inputs on the development reference and the embedded executable, checking numerical differences, memory use and timing. Quantisation, where the selected workflow uses it, should also be assessed against output quality. Successful conversion is one step, rather than validation of the whole application.
Firmware, model and signal preparation need to be versioned together. Changing sensor scaling, channel order or extracted features can make otherwise loadable weights inconsistent. Updates should check compatibility and integrity and provide recovery after interruption. Machine protection and functional safety require dedicated analysis.
How I would start an industrial product evaluation
I would start with measurements of the existing system and a precise diagnostic goal: which behaviour to recognise, with what maximum delay and acceptable false alarm rate. Then I would set a few verifiable prototype outcomes:
- Reliable acquisition: representative data, traceable sequences and handling of incomplete windows.
- Sustainable processing: measured memory use, a compatible model and a manageable window rate.
- Control validated under load: timing headroom and documented behaviour when processing is late or fails.
- Useful diagnostics: comparison against separate data and criteria understandable to the people using the alerts.
These findings allow STM32H7P to be compared with the current architecture. Combining functions on one MCU can be attractive, but removing an external component or reducing total cost depends on electronics, power, development and product validation.
Availability and where to start
As of 16 November 2026, ST reports sampling at OEMs and production planned for January 2027. NUCLEO-H7P5II and the STM32H7P5I-DK Discovery kit are announced for the same month. For a new project, board availability and lead times would therefore be among the first details to discuss with ST.
At Silicon LogiX, I work on microcontroller firmware, real-time architectures and on-device AI. To evaluate control and diagnostics on the same electronics, we can start with available signals, deadlines and the tests the prototype must pass.
Originally published on Silicon LogiX.



Top comments (0)